Speeding-up lattice sieving without increasing the
memory, using sub-quadratic nearest neighbor search
Anja Becker1 , Nicolas Gama2 , and Antoine Joux3,4
1
EPFL, Lausanne, Switzerland
2
UVSQ, Versailles, France
3
Laboratoire d’Informatique de Paris 6, UPMC Sorbonnes Universités, Paris
4
CryptoExperts, France and Chaire de Cryptologie de la Fondation de l’UPMC
anja.becker@epfl.ch, nicolas.gama@prism.uvsq.fr, antoine.joux@m4x.org
Abstract. We give a simple heuristic sieving algorithm for the m-dimensional
exact shortest vector problem (SVP) which runs in time 20.3112m+o(m) . Unlike
previous time-memory trade-offs, we do not increase the memory, which stays at
its bare minimum 20.2075m+o(m) . To achieve this complexity, we borrow a recent
tool from coding theory, known as nearest neighbor search for binary code words.
We simplify its analysis, and show that it can be adapted to solve this variant
of the fixed-radius nearest neighbor search problem: Given a list of exponentially
many unit vectors of Rm , and an angle γπ, find all pairs of vectors whose angle
≤ γπ. The complexity is sub-quadratic which leads to the improvement for lattice
sieves.
Keywords: Nearest neighbor search, lattice, sieve
1
Introduction
Lattice-based cryptography is a promising area due to the simple additive, parallelizable structure of a lattice. Additionally, lattices serve as a tool to attack
cryptosystems that are not necessarily lattice-based themselves. The two basic
hard problems shortest vector problem (SVP) and closest vector problem (CVP)
are known to be NP-hard1 to solve exactly [1, 16] and also NP-hard to approximate [11, 20] within at least constant factors. Moreover, hard lattice problems
have a long standing relationship to number theory and cryptology. In number theory, they can for example be used to find Diophantine approximations.
Together with the worst-case reduction by Ajtai [1] this provides a very good
starting point for lattice-based cryptography.
The time complexity of known algorithms that find the exact solution to
SVP or CVP are at least exponential in the dimension of the lattice. These
algorithms also serve as subroutines for strong polynomial time approximation
algorithms which have been shown to be of great use as a cryptanalytic tool. In
cryptology, they were used for a long time, first through a direct approach as
in [14] and then more indirectly using Coppersmith’s small roots algorithms [9,
10]. Algorithms for the exact problem enable us to choose appropriate parameters
for cryptographic schemes.
1
Under randomized reductions in the case of SVP.
A shortest vector can be found by enumeration [26, 15], sieving [2, 24, 22, 27,
6, 12, 28, 17, 18] or the Voronoi-cell algorithm [21]. Enumeration uses a negligible
2
amount of memory and its running time is between 2Õ(m) and 2Õ (m ) depending
on the amount and quality of the preprocessing. Probabilistic sieving algorithms,
as well as the deterministic Voronoi-cell algorithm are simply exponential in time
and memory.
Previous work has shown how to reduce the time complexity by exponential
factors while however increasing the memory requirement by exponential factors. As the memory requirement proves to represent the bottleneck in practice,
researchers are looking for only slight or no increase in the memory requirement.
Indeed, the most efficient sieve in practice is not the one of lowest asymptotic
time complexity. It is a parallelized version of the Gauss sieve [24, 7, 13, 23, 19] for
which the time complexity is unproven but postulated to be at most 20.41m+o(m) .
The complexity hence differs to less expensive sieves by exponential factors while
at the same time the memory complexity is smaller by an exponential factor. We
give details in the following paragraph and in Sect. 2. A provable variant of same
asymptotic complexity is the Nguyen-Vidick sieve [24] (NV sieve) for which we
propose a heuristic improvement.
Our contribution. The basic step in a lattice sieve algorithm is to search pairs of
vectors for which the norm of the sum or difference is smaller than the one of the
input vectors. The number of potential candidate vectors is exponential in the
lattice dimension and the cost to perform a reduction is hence also exponential.
It represents the dominating cost of the sieve and we are highly interested in a
most efficient execution.
Here, we propose a new and faster method that searches candidate vectors
for reduction. This reduces the currently used method which performs a trivial
exhaustive search over all vectors in the list. The new technique stems from a
new method for nearest neighbor search in the domain of coding theory [3]. We
develop a variant, presented in Sect. 3, that applies to the domain of lattices and
can replace the trivial reduction search in a lattice sieve by our method. The
complexity of an NV sieve [24] (cf. Sect. 2) is hence reduced from 20.41m+o(m)
to 20.3112m+o(m) without changing the required memory of 20.208m+o(m) in its
exponential factors5 as illustrated in Fig. 1. This stands in contrast to currently
proposed methods that reduce the time while increasing the memory by exponential factors.
Organization of the paper. In Section 2 we provide a brief background on Euclidean lattices and sieving algorithms (Sect. 2.1). We also summarize the nearest
neighbor search algorithm by May and Ozerov in Sect. 2.2 and discuss its limitations. The following Sect. 3 presents our new method to find neighbor vectors
which is based on the method by May and Ozerov. We provide an analysis of
the complexity and finally show how to apply it to the NV-sieve in Section 4.
5
The references in the figure refer to following publications in the bibliography: [NV08] =
[24], [MV10] = [21], [WLTB11] = [27], [BGJ14]= [6], [La14] = [17] and [LdW15] = [18]
20.45m
[NV08], [MV10]
11]
4]
LTB PH1 J14]
[W
[Z [BG
0. 40m
20.25m
20.20m
[La
1
J
[BG
14]
e=
sp
ac
e
4]
Ti
m
]
15
20.30m
nearest neighb our search
20.35m
dW
[L
Time Complexity
2
20.25m
20.35m
20.30m
Memory Complexity
20.45n
Fig. 1. New sieve vs. other time-memory trade-offs.
2
Required basic background
We here briefly summarize essential facts about Euclidean lattices, lattice problems for cryptographic use, important lattice algorithms and a nearest neighbor
search algorithm.
2.1
Euclidean lattices and lattice sieve algorithms
Euclidean lattices. A lattice L of dimension s ≤ m is a discrete subgroup of
Rm that spans an s-dimensional linear
Ps subspace. A lattice can be described as
the set of all integer combinations { i=1 αi bi | αi ∈ Z} of s linearly independent
vectors bi of Rs . Such vectors b1 , .., bm are called a basis of L. The
p volume of the
lattice L is the volume of span(L)/L, and can be computed as det(BB t ), for
any basis B. The volume and the dimension are invariant under change of basis.
Any lattice has a shortest non-zero vector of Euclidean length λ1 (L) which can
√
be upper bounded by Minkowski’s theorem as λ1 (L) ≤ m vol (L)1/m . The two
basic hard problems shortest vector problem (SVP) and closest vector problem
(CVP) are known to be NP-hard1 to solve exactly [1, 16] and also NP-hard to
approximate [11, 20] within at least constant factors. The problems can be solved
exactly or approximately by finding a shortest or closest vector up to a given
factor.
Sieving algorithms to solve SVP. Asymptotically, the most efficient method to
solve the exact SVP in practice is based on sieving. Lattice sieving algorithms
usually start with exponentially many large lattice vectors, and iteratively combine these vectors together, for instance replacing vectors by their difference
with other close vectors in the list, when they are shorter. The algorithm goes
on, while the norm of the vectors in the list decreases, until the list contains
1
Under randomized reductions in the case of SVP.
Algorithm 1 Nguyen-Vidick-like sieve
Input: A set L of vectors, of L, a parameter ε > 0
Output: A shortest vector of L
1: R ← maxNorm(L)
2: L0 ← ∅
3: repeat
4:
L0 ← {v ∈ L s.t. kvk ≤ (1 − ε)R}
5:
N ← FindNeighbors(L,γ = 31 − ε)
6:
for each (u, v) ∈ N do L0 ← L0 ∪ {u − v}
7:
L ← L0 , R ← (1 − ε).R
8: until |L0 | ≈ 1
9: return shortestVector in L
the shortest lattice vector, or at least, a very good approximation. Sieving algorithms differ by the strategy to find pairs of vectors close to each other in a
list of random lattice vectors. The simplest algorithms essentially test all pairs of
vectors using a quadratic-time approach to detect close vectors. For instance, the
NV-sieving algorithm [24] performs many straightforward iterations on the list,
each iteration computing at most one difference at a time, whilst the Gauss-sieve
algorithm [21] performs as many reductions as possible for each vector of the list,
and needs backtracking.
The sieving criterion to determine if the difference of two vectors (of approximately the same norm) is smaller is called a neighboring criterion, and is roughly
equivalent to testing whether the angle between the two vectors is ≤ π/3 − ε.
As explained in [24], a cone of angle π/3 covers a fraction sin(π/3)m of the
unit sphere, so a list of more than Õ(sin(π/3)−m ) = Õ(20.208m ) elements must
contain neighbor vectors. For this reason, all known sieving algorithms take as
input a list of 20.208m+o(m) vectors, and quadratic-time sieving algorithms have
complexity 20.415m+o(m) .
Various research has been done and is ongoing to reduce the complexity
of sieving algorithms [2, 6, 12, 22, 24, 27, 28]. Additionally, a new technique has
been developed recently that is based on locality-sensitive hash functions by
Charikar [8] and Andoni et al. [4, 5] which reduces the time and memory complexity in the asymptotic case down to 20.298m+o(m) [17, 18]. All these techniques
are time versus memory trade-offs, and therefore, increase the memory complexity by exponential factors. In practice, keeping the memory as small as possible
seems to be the most important factor to get a good running time, which is
confirmed by challenges [25], where quadratic algorithms with minimal memory
20.21m+o(m) achieve better than all time vs. memory trade-offs. For this reason,
we will focus on decreasing the running-time of sieving without affecting the
memory.
A simplified blueprint of Nguyen-Vidick-like lattice sieving is summarized
in Alg. 1, where the quadratic neighboring search phase has been extracted
as Alg. 2. Its asymptotic complexity is 20.41m+o(m) in time and 20.208m+o(m) in
memory for a random lattice of dimension m. In the next sections, we propose
faster alternatives for Alg. 2, which preserve the memory complexity.
Algorithm 2 FindNeighbors (QuadraticNNS, as in NV-sieve)
Input: List L = (u1 , . . . , uN ) of vectors, a parameter γ defining neighbors.
Output: A list of neighbor pairs (ui , uj ) in L2
1: N ← ∅
2: for i = 1 to N do
3:
for j = 1 to i − 1 do
4:
if angle(ui , uj )≤ πγ then N ← N ∪ (ui , uj ); break;
5:
end for
6: end for
7: return N
2.2
Nearest neighbor search for decoding by May and Ozerov, 2015
May and Ozerov recently presented an efficient method to find pairs of close
words in a lists of random binary words. They apply this algorithm to decode
binary linear codes [3]. Namely, given two lists L and R of |L| = |R| = 2λm
uniformly distributed and pairwise independent vectors from Fm
2 , find a pair in
L × R of small Hamming distance γm for γ < 1/2. The proposed algorithm
1
runs in time |L| 1−γ . The smaller γ, the better the running time. This stands in
contrast to previously known and widely used methods which are a subroutine
that perform a test on all tuples and hence run in time |L|2 . A similar task
appears also in lattice sieves for which we adapt the algorithm in Sect. 3.
The main idea of the algorithm from [3] is to successively apply randomized
filters on the input list, which may be computed individually on each word, and
which are designed to eliminate a large number of vectors. In their algorithm,
a filter corresponds to 2n random positions. A word is accepted by the filter if
and only if it contains at exactly n ones in these 2n positions, and exactly 1−δ
2 n
ones in the first n positions. δ ∈]0, 1[ and n ∈ N are optimization parameters
that they must tune to optimize the success probability.
The important point is that neighbor words have a much larger chance to
be accepted by the same filter than independent random words. Therefore, the
main idea is to apply different random filters to the input list of words, and to
recursively check whether neighbors are in the filtered sublists. If this process is
repeated with sufficiently many random filters, then each pair of neighbor words
is eventually be retrieved.
Discussion of limitations. In their algorithm, May and Ozerov [3] have to
deal with technical issues, which we remove in the case of lattices in Sect. 3
and Sect. 4. The first limitation is that the filtering condition is in practice very
restrictive. At each recursive level, the filter eliminates at least all words which
are unbalanced at the 2n positions. This yields large polynomial losses between
each iteration. These polynomial losses, raised to the power of the number of
levels, make their algorithm only usable for very large lists.
The second issue was that their binary words had a small length m. For
this reason, they had to choose m as their main complexity parameter, and at
each recursion level, the choices of n and δ had to be adapted to m. In practice,
n was proportional to m, and the proportion coefficient depends on γ. This
dependency between n, δ and m increases their number of recursion levels, which,
combined with the losses in the filtering conditions, make the overall algorithm
less practical.
Overall, the most important fact to remember is that asymptotically, when
the input words are very large, and the input list contains 2O(m) elements, then
in the most significant recursion levels, the parameter δ becomes very close to 0
and n is larger than 1/δ 2 . In this case, the overall complexity of the May-Ozerov
1
+ε
algorithm is equivalent to Õ(|L| 1−γ ), which is sub-quadratic, and gets smaller
when γ decreases.
3
Nearest neighbor search adapted for sieving
We here describe a variant of [3] as described in Sect. 2.2 which we will use
as a subroutine in a sieving algorithm to find vectors suitable for reduction as
we present in Sect. 4. Our algorithm overcomes the limitations of the previous
nearest neighbor search algorithm as discussed in the above section. In contrast
to the May-Ozerov proposition, we ensure that the filtering conditions are easier
to fulfill, and we obtain non-negligible filtering probabilities. We show that in
our nearest neighbor search setting, the parameters (equivalent to n and δ) can
be arbitrarily fixed, which simplifies a lot the analysis. The overall complexity is
1
+ε
again of order Õ(|L| 1−γ ) searching neighbors at an angle γπ in an input list L.
It hence improves over a classic search of pairs of vectors performed in quadratic
time.
Setting. We are given a list of N uniformly distributed real vectors of Rm of norm
one, and we wish to find all pairs of vectors of an angle inferior to γπ (radians),
that we call neighbors. If the input list has N elements, we expect to find about
N 2 sin(γπ)m pairs of neighbors. The naive approach is the quadratic nearest
neighbor search (QuadraticNNS), which tests all pairs, and which is essentially
optimal when N is arbitrarily large. Instead, we will focus on the particular
case where N is smaller than 1/ sin(γπ)m . In this case, we expect to find O(N )
neighbors, and we will adapt and simplify the algorithm of [3] to find these pairs
1
+ε
in sub-quadratic time O(N 1−γ ).
Basic principles. The main idea of the algorithm remains to successively apply
randomized filters, which are efficient functions from Rm to {0, 1}. Random vectors should be accepted by the filter with some low probability Pf . But most
importantly, a pair of neighboring vectors must have a much larger probability
Pp to be simultaneously accepted by the filter than Pf2 for a pair of randomly
chosen vectors. As we will see later, this difference of probability between Pp and
Pf2 is precisely what causes the sub-quadratic complexity.
In our setting, the set of random filters is denoted F(n, δ), and is parametrized
by a number n ∈ N of tests, and a weight constraint δ ∈]0, 1[. To instantiate a
filter f ∈ F, we choose n ∈ N>0 directions ui uniformly over the 1-sphere of Rm .
A non-zero vector v ∈ Rn passes the filter f if and only if there are at most 1−δ
2 n
directions ui such that hv, ui i ≥ 0. In this case, we set f (v) = 1, else f (v) = 0
means that the vector v is rejected.
Procedure. Let γ < 1/2 be fixed (γπ is the angle between neighbors) and let
δ > 0, n ∈ N and K be three optimization parameters. Let L be a list of N real
vectors in Rm , whose directions are uniformly distributed on the sphere of radius
one. The goal is to establish the list of all neighboring pairs.
– If N is too small (say N ≤ Nmin ), we use the naive quadratic algorithm,
which checks all pairs.
– If N is larger, we repeat K times the following: We choose a random filter
f ← F(n, δ), we extract only the sublist L0 = {v ∈ L s.t. f (v) = 1} of
vectors which pass the filter and we continue recursively on L0 .
Algorithm 3 presents a pseudo code of the algorithm.
Algorithm 3 Nearest neighbor search (NNS)
Require: A list L of N unit vectors of Rm
Ensure: The list of all neighboring vectors
Global: Repetitions K, an angle πγ defining neighbors, filtering parameters δ, n, a lower-bound
Nmin ∈ N for the recursion
1:
2:
3:
4:
5:
6:
7:
8:
9:
10:
if N < Nmin then
Return (v, w) ∈ L2 s.t. hv, wi ≥ cos(πγ) via QuadraticNNS(L)
end if
N ←∅
for i = 1 to K do
Pick a random filter f ∈ F
Filter L0 = {v ∈ L s.t. f (v) = 1}
N ← N ∪ NNS(L0 )
end for
Return N
Analysis. The success of Alg. 3 depends on two probabilities. We call Pf (δ, n)
the probability that a vector v is accepted by a random filter, and Pp (γ, δ, n)
the probability that two vectors v, w, having an angle γπ, are simultaneously
accepted by the same random filter. Due to the spherical symmetry of the problem, we may arbitrarily fix v = (1, 0, . . . , 0) and w = (cos(γπ), sin(γπ), 0, . . . , 0)
without affecting the probabilities when calculating them.
The first probability is easy to estimate: v is accepted if and only if at most
j0 = 1−δ
2 n directions are in the same hemisphere as v which happens with
probability
1−δ
n 2
1 X
n
Pf (δ, n) = n
.
(1)
2
j
j=0
Since the vectors in the input list are uniformly distributed on the sphere, we
thus expect to keep N · Pf elements in L0 on average after the filtering step on N
elements. The number of accepted elements decreases quickly when δ increases,
as the weight requirement on the vectors becomes more restrictive.
Heuristic 1 (Number of elements in lists). We assume that in each step of
Alg. 3 we filter N Pf elements given N elements.
The second probability Pp is a bit longer to evaluate. We analyze the setting
as illustrated in Fig. 3 where we see two vectors at an angle γπ and the intersection of the two hemispheres Hv , Hw defined by v, w. We draw n directions
ui and observe in which part of the intersecting hemispheres they fall. We now
describe the dependency between: (i) the angle and (ii) the number of directions
that fall in the four sections defined by the hemispheres. Let us assume the following scenario: At most j0 directions are in the hemisphere Hv centered in v,
and at the same time, at most j0 directions are in the hemisphere Hw centered
in w. If we note i the number of vectors in Hv ∩ Hw , j the number of vectors
in Hv ∩ H w , k the number of vectors in H v ∩ Hw , the exact value of Pp can be
computed as:
n!
1 X
1−δ
Pp (γ, δ, n) = n
γ j+k (1 − γ)n−j−k with j0 =
n. (2)
2
i!j!k!(n-i-j-k)!
2
i∈[0,j0 ],
j,k∈[0,j0 −i]
By definition of the probability Pp , it suffices to repeat the main loop about
K = ln(x) · Pp−1 times, so that with probability ≥ 1 − x1 , each pair of neighbors
appears at least once in the same sublist. As long there are at most x recursive
levels in total, an elementary induction on the recursive calls in Alg. 3 proves
that each pair of neighbors is recovered with constant probability.
Exponent of complexity
v
w
1−γ
2
γ
2
γπ
1.72
1.7
1.68
1.66
1.64
1.62
1.6
1.58
1.56
1.54
1.72
1.7
1.68
1.66
1.64
1.62
1.6
1.58
1.56
1.54
i dirs
j dirs.
0.9
0.7
10
γ
2
20
30
0.5
40
50
60
n
70
0.3
80
90
100
δ
0.1
k dirs.
1−γ
2
n-i-j-k dirs
Fig. 2. Intersection of hemispheres.
3.1
Fig. 3. Development of exponent for different choices of n, δ, γ = 1/3.
Complexity analysis for fixed parameters
In this section, we analyze the complexity of our nearest neighbor search
(cf. Alg. 3) when the three parameters γ, δ and n are fixed. This means that
the probabilities Pp and Pf are constant and we can compute the complexity
as presented in the following. We also provide example parameters to illustrate
their size and impact on the complexity. We now state our main contribution of
the section. Given as input a list of N words, we note C(N ) the time complexity,
expressed as a number of dot products on Rm .
Theorem 1 (Complexity of NNS). Let γ < 1/2, δ < 1 and n be fixed. Let
L be a list of N ≤ poly(m) sin(πγ)−m uniformly random vectors in the sphere of
dimension m. We call d = ln(Pp−1 )/ ln(Pf−1 ) the exponent such that Pfd = Pp .
Then the time complexity of Alg. 3 is Õ(N d ).
Proof. The overall time complexity to treat a list of length N is given by
1
C(N ) = N 2 when N ≤ Nmin
2
C(N ) = 1/Pp · (C(Pf N ) + nN )) when N > Nmin
(3)
(4)
The first row corresponds to the quadratic algorithm. The second row includes the 1/Pp repetitions in which we count one recursive call and one filtering
process, composed of n dot products per vector. The induction goes on until the
list is very small, which typically means, Nmin = poly(m). In order to simplify
and solve the above recurrence, we set X = 1/Pf and uk = C(Nmin X k ) for
k ∈ N. Then, the sequence (uk )k∈N satisfies the following linear recurrence:
1 2
u0 = Nmin
, uk = X d uk−1 + Nmin X d+k
2
which has the explicit solution:
!
1−d d
X
Nmin
1 2−d
Xd
Nmin + d−1
(Nmin X k )d − d−1
(Nmin X k )
uk =
2
X
−1
X
−1
Thus, assuming that the complexity of the algorithm increases with N , the final
asymptotic complexity satisfies for all N ≥ Nmin
1−d −1
2−d
Pf )N d .
+ 2Nmin
C(N ) ≤ (Nmin
We recall that this algorithm is particularly interesting for N =
Õ(sin(πγ)−m ) which is exponential in the dimension of the vector space. Therefore, it suffices to choose Nmin polynomial or slightly sub exponential in m so that
the whole factor in C(N ) vanishes before N d . Therefore, we have proved that
each choice of parameters n, δ, γ corresponds to an exponent d such that asymptotically, finding all pairs of neighbors in a list of size N ≤ poly(m)/ sin(γπ)m
can be achieved with time complexity Õ(N d ), and memory O(N ).
To get a feeling of the parameters in practice, we give some example parameters n, δ in Table 1. We fix γ = 1/3 and compute the resulting exponent d of
the complexity. Also Fig. 3 illustrates the complexity development as n and δ
change, again for γ = 1/3.
The next section analyses the asymptotic case and presents the value to which
d tends. The analysis shows that asymptotically 1.5 + ε for γ = 1/3 and small
ε < 1 is achieved. The development in Tabel 1 gives us a glance of this behavior
in moderate dimensions.
3.2
Complexity analysis in the asymptotic case
In this section, we prove the following theorem, which proves that for each angle
γπ, there exists a choice of parameters n, δ for which the complexity exponent
1
of our nearest neighbor search decreases to 1−γ
.
Table 1. Practical values of n, δ for γ = 1/3.
n
30
50
70
90
30
40
50
60
70
90
150
200
δ
0.15
0.25
0.25
0.25
0.4
0.4
0.4
0.4
0.4
0.4
0.4
0.35
d
1.6491
1.6088
1.6026
1.5910
1.6019
1.5919
1.5842
1.5781
1.5730
1.5652
1.5507
1.5493
1/Pf (≈ Nmin )
5.53
30.81
48.32
134.53
46.76
120.56
303.01
748.94
1829.42
10652.1
1.88883e+06
2.40156e+06
1/Pp (≈ K)
16.79
248.32
499.97
2436.99
473.13
2056.17
8534.24
34366.4
135444
2.01164e+06
5.39998e+09
7.68068e+09
Theorem 2. Let γ ∈]0, 21 [ be a parameter and ε > 0 be a small quantity, there exists parameters n ∈ N and δ > 0 such that the exponent d =
1
ln(Pp (γ, n, δ))/ ln(Pf (n, δ)) is ≤ 1−γ
+ ε. For these choices of parameters, the
time complexity of Alg. 3 on an input list of N = sin(πγ)−m vectors is asymp1
+ε
totically Õ(N 1−γ ).
To prove this theorem, we compute asymptotic equivalents of ln(Pp ) and
ln(Pf ) for large n ∈ N, and small δ > 0.
Proof. The probability Pf in (1) is a sum of binomial terms, and since the last
index j0 = 1−δ
2 n is lower than n/2, Pf satisfies the following bounds:
1
2n
n
1−δ
2 n
≤ Pf (n, δ) ≤ n ·
1
2n
n
1−δ
2 n
.
Using the Stirling formula, and the entropy function H(x) = x log2 (x) + (1 −
x) log2 (1 − x), we deduce that asymptotically,
ln(Pf−1 ) ∼
n→∞
1−δ
n ln(2) 1 − H
2
(5)
For small values of δ, the entropy function has the following Taylor expansion:
2
3
= ln(2) − ln(2)
2 δ + ◦(δ ). Therefore, we also have the simpler equivalent:
H( 1−δ
2 )
ln(Pf−1 )
∼
δ→0
nδ 2 →∞
1 2
nδ
2
(6)
We now do the same for the probability Pp , and start by finding the maximal
term in the expression of Pp from (2). First, we remark that the multinomial
n!
coefficient is the product between i!(j+k)!(n−i−j−k)!
and the binomial coefficient
(j+k)!
j!k! .
When the sum j +k is fixed, this binomial coefficient reaches its maximum
when j = k. Then, writing i = ni0 and j = nj 0 , the maximal term of the
expression of Pp corresponds to the maximum of the function:
0
0
f (i , j ) := ln
n!
0
γ 2nj (1 − γ)n−2nj
0
0
2
0
0
ni !(nj !) (n − ni − 2nj )!
1−δ
0
0
over the domain i0 ∈ [0, 1−δ
2 ], j ∈ [0, 2 − i ]. We may approximate
with its equivalent via Stirling formula:
1
0 0
n f (i , j )
f (i0 , j 0 )
≈ −i0 ln(i0 )−2j 0 ln(j 0 )−(1−i0 −2j 0 ) ln(1−i0 −2j 0 )+2j 0 ln(γ)+(1−2j 0 ) ln(1−γ)
n
A quick computation of the partial derivatives shows that this function increases with i0 and j 0 on its entire domain. The maximum is reached at the
0
0
upper-frontier of the domain where i0 = 1−δ
2 − j . Then, substituting i with
1−δ
0
0
0
2 − j and derivating again, we get that the extremum is reached for j = j0
satisfying:
ln(
1+δ
1−γ
1−δ
− j00 ) − 2 ln(j00 ) + ln(
− j00 ) − 2 ln(
)=0
2
2
γ
or equivalently, j00 is the positive solution of the second degree equation:
"
1−γ
γ
#
2
−1
2
.j00
+
j00
−
1 − δ2
4
=0
This is the exact expression of j00 .
j00
γ
=
2
−γ +
!
p
(1 − γ)2 − (1 − 2γ)δ 2
1 − 2γ
(7)
Finally, ln(Pp (γ, n, δ)) admits the following equivalent when n grows:
1−δ
0 0
ln(Pp (γ, n, δ)) ∼ n · ln(2) − f
− j0 , j0
2
(8)
γ
Again, computing the Taylor expansions of j00 = γ2 − 4(1−γ)
δ 2 +◦(δ 2 ) from (7),
0 0
and then substituting in f 1−δ
2 − j0 , j0 , we obtain a simpler equivalent when δ
2
is small and n is larger than 1/δ .
ln(Pp (γ, n, δ))
∼
δ→0
nδ 2 →∞
nδ 2
2(1 − γ)
(9)
Finally, dividing the two equivalents (9) and (6), we obtain the limit exponent, which proves the theorem:
d=
ln(Pp (γ, n, δ))
ln(Pf (n, δ))
∼
δ→0
nδ 2 →∞
1
1−γ
(10)
4
Application to lattice sieves
We can apply the above algorithm to solve a shortest vector problem (SVP)
or approximate SVP for a given lattice L of dimension m by means of a sieving
algorithm. For instance, the NV sieving algorithms first samples an exponentially
large set L0 , of N = 20.2075m+o(m) long lattice vectors of same norm R0 . Then,
a sieving iteration step selects pairs of neighbor vectors from L0 × L0 whose
distance is ≤ (1 − ε)R0 , stores their differences in a new list L1 , and continues
with L1 .
The NV strategy reduces the running-time by a polynomial factor by ensuring
that a vector of the list can be reduced by at most one other vector. If this
approach is faster than testing all pairs, it also has the drawback to decrease the
number of returned neighbor pairs for L1 , which is by definition always smaller
than L0 . And overall, the complexity of an iteration remains quadratic in N .
Here, the idea is to replace a sieving iteration with our sub-quadratic nearest
neighbor search algorithm. Since two neighbors have an angle smaller than π/3,
which corresponds to γ = 31 , this can effectively decrease the time complexity
from Õ(N 2 ) to Õ(N 1.5+ε ), depending on the choices of n, δ as in the previous
section. More precisely, the asymptotic time complexity drops from 20.4150m+◦(m)
to 20.3112m+◦(m) , and the memory remains of the order of the input 20.2075m+◦(m) .
The substitution of the nearest neighbor search in sieving algorithms is
straightforward, modulo the usual minor differences in the distribution of vectors
that need to be taken into account.
– First, the input vectors have a discrete distribution on the lattice instead of
a continuous radial distribution over Rm . In the first iterations of sieving, it
is likely that the directions of Gaussian lattice vectors are indistinguishable
from random directions in Rm . However, since the radius of the pool of vectors
decreases by a constant factor 1 + ε at each iteration, it will eventually reach
a critical radius where there are not enough lattice vectors to fill the input
list. This was already the case in the original NV-sieve, and it is the main
reason why sieving ends up with the shortest lattice vector.
– In the end of each iteration, the difference between each pair of neighbors
is put into the input list for the next iteration. At this point, we need to
use the heuristic that the difference of all neighbor pairs have uniformly
distributed directions in Rm . This is actually the main heuristic in every
sieving algorithms.
5
Conclusion
We have presented a new nearest neighbor search algorithm which enumerates
all pairs of neighbor vectors in the unit sphere of Rm . This has the immediate application to reduce the complexity of lattice sieving without affecting the
memory. It remains open to know where is the practical cross-over point between
this strategy and other sieving techniques.
References
1. M. Ajtai. The shortest vector problem in L2 is NP-hard for randomized reductions (extended abstract). In Proceedings of the 30th annual ACM symposium on Theory of computing, STOC ’98, pages 10–19, New York, NY, USA, 1998. ACM Press.
2. M. Ajtai, R. Kumar, and D. Sivakumar. A sieve algorithm for the shortest lattice vector
problem. In Proceedings of the 33rd annual ACM symposium on Theory of computing,
STOC 2001, pages 601–610, New York, NY, USA, 2001. ACM Press.
3. I. O. Alexander May. On computing nearest neighbors with applications to decoding
of binary linear codes. In Advances in Cryptology – Eurocrypt 2015, Lecture Notes in
Computer Science. Springer, 2015.
4. A. Andoni, P. Indyk, H. L. Nguyen, and I. Razenshteyn. Beyond locality-sensitive hashing.
In SODA, pages 1018–1028, 2014.
5. A. Andoni and I. Razenshteyn. Optimal data-dependent hashing for approximate near
neighbors. In SODA, 2015.
6. A. Becker, N. Gama, and A. Joux. A sieve algorithm based on overlattices. LMS Journal
of Computation and Mathematics, 17:49–70, 2014.
7. J. W. Bos, M. Naehrig, and J. van de Pol. Sieving for shortest vectors in ideal lattices: a
practical perspective. Cryptology ePrint Archive, Report 2014/880, 2014.
8. M. S. Charikar. Similarity estimation techniques from rounding algorithms. In STOC,
pages 380–388, 2002.
9. D. Coppersmith. Finding a small root of a bivariate integer equation; factoring with high
bits known. In U. Maurer, editor, Advances in Cryptology – EUROCRYPT 1996, volume
1070 of Lecture Notes in Computer Science, pages 178–189. Springer Berlin Heidelberg,
1996.
10. D. Coppersmith. Finding a small root of a univariate modular equation. In M. Darnell,
editor, Advances in Cryptology – EUROCRYPT 1996, volume 1070 of Lecture Notes in
Computer Science, pages 155–165. Springer Berlin Heidelberg, 1996.
11. I. Dinur, G. Kindler, and S. Safra. Approximating-cvp to within almost-polynomial factors
is np-hard. In Proceedings of the 39th Annual Symposium on Foundations of Computer
Science (FOCS ’98), pages 99–109, Washington, DC, USA, 1998. IEEE Computer Society.
12. G. Hanrot, X. Pujol, and D. Stehlé. Algorithms for the shortest and closest lattice vector problems. In Coding and Cryptology - Third International Workshop, IWCC 2011,
Qingdao, China, May 30-June 3, 2011. Proceedings, pages 159–190, 2011.
13. T. Ishiguro, S. Kiyomoto, Y. Miyake, and T. Takagi. Parallel Gauss sieve algorithm :
Solving the SVP in the ideal lattice of 128-dimensions. Cryptology ePrint Archive, Report
2013/388, 2013.
14. A. Joux and J. Stern. Lattice reduction: A toolbox for the cryptanalyst. Journal of
Cryptology, 11(3):161–185, 1998.
15. R. Kannan. Improved algorithms for integer programming and related lattice problems. In
Proceedings of the 15th annual ACM symposium on Theory of computing, STOC ’83, pages
193–206, New York, NY, USA, 1983. ACM Press.
16. R. M. Karp. Reducibility among combinatorial problems. In R. E. Miller and J. W.
Thatcher, editors, Complexity of Computer Computations, The IBM Research Symposia
Series, pages 85–103. Plenum Press, New York, 1972.
17. T. Laarhoven. Sieving for shortest vectors in lattices using angular locality-sensitive hashing. IACR Cryptology ePrint Archive, 2014:744, 2014.
18. T. Laarhoven and B. de Weger. Faster sieving for shortest lattice vectors using spherical
locality-sensitive hashing. IACR Cryptology ePrint Archive, 2015:211, 2015.
19. A. Mariano, S. Timnat, and C. Bischof. Lock-free GaussSieve for linear speedups in parallel
high performance SVP calculation. SBAC-PAD’14 - 26th International Symposium on
Computer Architecture and High Performance Computing, 2014.
20. D. Micciancio. The shortest vector in a lattice is hard to approximate to within some
constant. In Proceedings of the 39th Annual Symposium on Foundations of Computer
Science, FOCS ’98, pages 92–, Washington, DC, USA, 1998. IEEE Computer Society.
21. D. Micciancio and P. Voulgaris. A deterministic single exponential time algorithm for most
lattice problems based on Voronoi cell computations. In Proceedings of the 42nd ACM
22.
23.
24.
25.
26.
27.
28.
symposium on Theory of computing, STOC ’10, pages 351–358, New York, NY, USA, 2010.
ACM.
D. Micciancio and P. Voulgaris. Faster exponential time algorithms for the shortest vector
problem. In Proceedings of the twenty-first annual ACM-SIAM symposium on Discrete
Algorithms (SODA ’10), pages 1468–1480. ACM/SIAM, Society for Industrial and Applied
Mathematics, 2010.
B. Milde and M. Schneider. A parallel implementation of gausssieve for the shortest vector
problem in lattices. In Parallel Computing Technologies, volume 6873 of Lecture Notes in
Computer Science, pages 452–458. Springer Berlin Heidelberg, 2011.
P. Q. Nguyen and T. Vidick. Sieve algorithms for the shortest vector problem are practical.
Journal of Mathematical Cryptology, 2(2):181–207, 2008.
M. Schneider, N. Gama, P. Baumann, and P. Nobach. http://www.latticechallenge.org/svpchallenge/halloffame.php.
K.-P. Schnorr and M. Euchner. Lattice basis reduction: Improved practical algorithms and
solving subset sum problems. Mathematical Programming, 66:181–199, 1994.
X. Wang, M. Liu, C. Tian, and J. Bi. Improved Nguyen-Vidick heuristic sieve algorithm
for shortest vector problem. In Proceedings of the 6th ACM Symposium on Information,
Computer and Communications Security, ASIACCS ’11, pages 1–9, New York, NY, USA,
2011. ACM.
F. Zhang, Y. Pan, and G. Hu. A Three-Level Sieve Algorithm for the Shortest Vector
Problem. In T. Lange, K. Lauter, and P. Lisonek, editors, SAC 2013 - 20th International
Conference on Selected Areas in Cryptography, Lecture Notes in Computer Science, pages
29–47, Burnaby, Canada, 2013. Springer Berlin Heidelberg.